0:00Arun Kalaiselvan, cofounder and CTO of Arya, says it is a workforce management platform for in home health care covering scheduling agents, onboarding, credentials and compliance, and its AI phones caregivers with a bonus offer to fill open shifts.
2:23Sam Nadler demos a prototype that turns a CHF discharge summary into a start of care packet with risk factors, a four week schedule and matched staff. Anand Chandrasekaran says it captured the workflows well, but real intake is not real time.
7:27Sam Nadler says he researched Arya Health, got a PRD from ChatGPT and built the demo in Google AI Studio with Gemini Pro in about thirty minutes and four or five tweaks, but real world use takes many more steps.
8:59Arun Kalaiselvan says about 70% of Arya's infrastructure is self hosted, vendors sign HIPAA specific agreements, customer data is separated, and AI is used only through cloud providers: Anthropic via Amazon Bedrock and Gemini via Google Vertex AI.
13:51Arun Kalaiselvan says an Anthropic connection change made one bill six times higher than expected. Anand Chandrasekaran says newer models turn many multi shot tasks into single shot, allow more precise prompts, and Gemini 3 and 2.5 Flash are very fast.
18:08Anand Chandrasekaran says Pydantic AI's dependencies let typed inputs steer an agent's thinking, so 90 to 95% of everyday agent tasks need no examples, and his hallucinations dropped 40 to 50% after moving off JSON types.
19:44Arun Kalaiselvan says finding a replacement can take an admin two to three hours; Arya knows who is available, credentialed and suited to the patient and calls them. One customer had about 12 pediatric shifts filled over a weekend.
23:20Arun Kalaiselvan says Arya now builds care teams around patients instead of filling shifts. Anand Chandrasekaran says a heuristic eligibility system scores caregivers in tiers, and Arun says they are studying drive time as a traveling salesman problem.
27:29Arun Kalaiselvan says AI dating may help the first meet, but AI only knows your highlights, and the winners will capture more than that. Anand Chandrasekaran asks how true AI-written dating profiles will be.
31:08Jordan Metzner says Cursor is his go to IDE but model makers may build the best tooling. Anand Chandrasekaran expects coding models to saturate, making tooling and price decisive. Arun Kalaiselvan calls Google the silent killer.
AI Scheduling for Home Health: Filling Shifts in Minutes: the key questions from this episode, answered with quotes and timestamps.
Arun Kalaiselvan [0:00]: We literally have AI give them a phone call and say, hey. This patient's shift opened up today. They need help. We'll give you a $10.20 dollar bonus to go apply to do the shift and get it filled instantly right there.
[0:16] Theme song: Built this week, breaking it down. Built this week, we show you how. A fresh idea, a clever tweak, you locked in shoe. Built this week.
Sam Nadler [0:25]: Hey, everyone, and welcome to Built this week, the podcast where we share what we're building, how we're building it, and what it means for the world of AI and startups. I'm Sam Nadler, cofounder here at Ryz Labs. And each and every week, I'm joined by my friend and business partner cohost, Jordan Metzner. And this week, we have two special guests from Arya Health, Arun and Anand. And, Arya, I'm gonna let them introduce themselves, but it's a platform for home health agencies that uses AI to help hospital handoffs, start of care, clinician scheduling, and just really helps improve efficiencies and do it faster. But Arun and Anand, welcome to the podcast. Please give me a quick intro about yourselves and Arya Health.
Arun Kalaiselvan [1:14]: Good day. Thank you for having us. My name is Arun, cofounder and CTO at Arya. Arya is a workforce management platform currently focusing on the in home health care space, like you said. So we have everything from scheduling agents to onboarding employees, managing their credentials, managing compliance, things that you don't normally think about that office managers do. And then it's a lot of mundane work that they do, and we take all of that away and automate it so that they can focus on revenue generating work instead. And, Anand?
Anand Chandrasekaran [1:50]: Hi, guys. Thank you for having me as well. This is Anand. I'm the principal architect here at Arya Works. Like Arun was saying, we actually use AI for doing all the mundane work that people would do in health care space, especially in home health. We try to do that in a really secure and safe way so that they don't have to worry about, like, any security concerns. That's actually one of our most important aspects. Like, we really care for what they do. At the same time, we wanna make sure that it's all secure.
Sam Nadler [2:22]: Amazing. Well, just to jump in, we built something with Arya in mind. And in our call, we discussed potentially building something for the start of care intake. And I'm not from the healthcare space, so this could be a little bit off, but I wanted to create a few different demos. Our first one is a complex CHF. So you would get something like this discharge summary from the hospital. Again, I don't know if this is what it actually looks like. Let's use AI to generate the start of care packet that would help inform from this, I would say, not messy, but maybe unstructured and somewhat coded discharge summary. So as you go through, you can see in this discharge plan, it, has this two w three, then one w four. Maybe for a health care audience, that that's obvious, but it's not so obvious for me. They have certain, home situations, for home care, whether they have, pets or stairs or they live alone. I can imagine all that could be important to understand who's the right person to help with this patient. But anyway, we trigger the start of care packet, and as you can see, it gives, exactly what's happening, congestive heart failure, risk factors. So you have the recent c h CHF exasperation. You have the clinical strategy.
Sam Nadler [3:52]: I thought this little projected four week schedule was cool. You have the diffs the different disciplines, how often they need to be there in the first week. So week one, the s in, I'm not even sure what that stands for, but I guess some sort of wound care nurse would need to be there every day for the first week, five days for the second week, etcetera. And you can see kind of what that schedule immediately matches theoretical staff. So you have Sarah Chin, oh, skilled nurse, that makes sense, who's, has wound care competencies. She speaks Spanish. Wow. Let's go ahead and send Sarah a auto drafted outreach to see if she can join, and this obviously doesn't work, but that's what we would do. It kinda drafts these, and you could theoretically engage with these different medical personnel quickly. Start of care details, you have the address, you have the level of urgency, some home contacts, which I mentioned before.
Sam Nadler [4:54]: They have four steps to enter home, elderly wife, at least someone's available, but maybe she it's will be unable to lift the patients. You got some missing info, gender's missing, maybe, one of the teammates of Arya would follow-up with this, maybe some are more important than others, and then you have some workflow actions that you could just go ahead and click through as you get through this patient's start of care. And then lastly, as just a little fun feature, I created a get well card functionality where if we click it and it's supposed to be punny and funny. Maybe that's not the right, moment to send, but you got that get well soon, Arthur, from Arya Health. Arthur sending you heartfelt wishes for a speedy recovery. We're rooting for you to feel hearty and strong, ready to take those steps back home soon. So this is The start of care packet. Again, I'm not from the health care space. Maybe it's off base. Maybe it's somewhat aligned, but would love to hear your thoughts.
Arun Kalaiselvan [6:02]: Well, I love that you and Medieval are on this, first off. That's fantastic. I actually think this is a pretty good starting point for the start of care workflow. Anand does I haven't seen Anand nod along so much So to new products.
Anand Chandrasekaran [6:19]: It's interesting that you captured some of the workflows really well. But what's most interesting in this is something that we are hoping to do in, the fourth version is already captured in this. Yeah. And I really like that part about it. Like so I mean, it's it's it's difficult to get into the SOC stuff and then be like, you cannot just, like, own the process immediately. The eventual goal is to own the process, right, from the first. Right? But, yeah, this captures that very well. I really like that. There is a lot of things that happens in this that takes we wish it is real time, but it is not real time. That's the only thing I think is off. That's all. I think I really like this kind of a summary that, like, this is cool and I really like it.
Arun Kalaiselvan [7:03]: Yeah. I definitely want to Yeah. Yeah, give shoot us a copy. It got all the titles right. That was cool. I The skilled nursing part, the occupational therapy, the physical therapy, and the home health aide. It got some details pretty spot on.
Anand Chandrasekaran [7:17]: I wouldn't believe if you tell me that you're not from the health care space, I wouldn't believe you from this. If I see this, I would be like, yeah. You got it right. So that's that's nice.
Jordan Metzner [7:27]: That's cool. Well, Sam, tell them your workflow. Like, I mean, I think that that'd be cool because, I mean, you entered a work you know, we talk about AI. You don't really this space that well. It'd be cool to talk about your workflow just a little bit on, like, how you build something like this without your domain expertise.
Sam Nadler [7:40]: Yeah. I mean, it's not you know, I'm a little embarrassed how simple it was, but basically did a little research on Arya Health with the context that we were looking to build a feature for start of care. Got a PRD from ChatGPT, went to the Google AI studio using Gemini Pro, put the PRD, and then within four or five different tweaks, probably in the span of thirty minutes, I had kind of what we're seeing now. And then I just went just small refinements to add additional features, but at the end of the day, it was really fast from just a conceptual perspective. And obviously, that's what we see is AI is so powerful, is you can go from, like, zero to one so fast. Now, actually, in as you guys and I would love to better understand how Arya is is working today. But, a, taking this from a concept to actual real world use cases is there's a million steps in between. Like, it's not in real time, as you mentioned. The discharge notes, I don't even know how those come in. Maybe it's phone call.
Sam Nadler [8:57]: Yeah. I don't even know.
Jordan Metzner [8:58]: Well, let's get into the good stuff. It's all about, I think, as Anand mentioned, about compliance, right, and patient privacy and HIPAA. Right? And so, you can't just, push this thing to some firebase or some supabase or some amplify project and run away with it like you can on some, other vibe coded projects. So I think, this is probably a good time where we talk to you guys about, obviously, there's a lot of model development. There's a bunch of new models coming out on a weekly basis from a coding perspective. Obviously, like, we're deep on all of them. I mean, obviously, Opus 4.5 has just been insane so far. But, I wanna talk to you guys about how you use AI in a privacy forward focus where you obviously, some of these new hosted models come out, next week. But, how easy is it to be able to adopt those models in a privacy, environment where patient, privacy is like number one?
Arun Kalaiselvan [9:52]: Yeah. It's I think that is the most complicated bit of this whole ecosystem. It's it's not as simple as just, like, hopping on chat.openai and starting talk starting the conversation with ChatGPT. Right? So one of the things that we care about deeply at Arya is security and compliance. So much so that pretty much I wanna say about 70% of our infrastructure is completely hosted by us. And then the bits that's not hosted by us is primarily the communication pieces, which is our phone calls, SMSs, and that kind of stuff, which we have direct sort of agreements with our vendors. So if we work with Bland, if we work with Twilio, we work with Amazon's Pinpoint, we have agreements in place that are completely, specific to HIPAA regulations, and that sort of covers most of our use cases. But then we go beyond that. Every customer gets almost a physically separate location where their data is stored and there's definitely logical separations among everywhere.
Arun Kalaiselvan [10:59]: There is never a case where any data will ever leak from one customer into another customer's ecosystem. At Arya and speaking about AI, we only use AI through our cloud providers. So we work with Amazon and Google. So with Amazon, we work with Bedrock. We talk to Anthropic through Bedrock. And then on Google, we work with the Vertex AI, and we talk to Gemini through Vertex AI. And they make it possible for us to build HIPAA compliant workflows just because that we have those contractual relationships with Amazon and Google, and then they cover all of the all the data at rest as well as in transit, and they cover that under our security policies. The cool thing about Google is Anand actually advises them on how to build their agent development kits. They're on our Slack. We have conversations pretty much almost every day, and they're you know, we are huge fans. They're huge fans. It's good. It's been fun.
Anand Chandrasekaran [12:02]: Yeah. I wouldn't call it advice, but yeah. Some words like, we Share not. Yeah. I can talk a Little bit about that. So so the thing is, if you look at it, right, like AI and security, they are still not going hand on hand. There are so many tools available, but if you see which of them is HIPAA compliant to the extent of to be used by PHA, there's it's very limited. So from the first, our model has been, like, you can only use the models that's available in our cloud provider because only that is giving us the utmost security. That means that, like, we will experiment quite a lot within what's available in that. Personally, what my process right now is I check the models that's available in cloud and other things constantly on a site. Whereas for our well, Arya workflows, we only use those that is available in cloud. And, like, sometimes we push them to make them available. I remember Bedrock, like, initially, they didn't have it. Like, we pushed them to get a particular version.
Anand Chandrasekaran [13:01]: A lot of them did, I guess, so we got that. With Google, we have great support. So if we need something, are more than willing to do that. And we have that relationship with them, and it's really worked out very well for us. One more point I wanna add, we are multi cloud. But even in multi cloud, the way we communicate between our clouds, we have a separate tunnel communication happening. Even in that level, the security is extremely yeah. I mean, it's not paranoia, but at the same time, we wanna make sure that security comes first. So we have a separate peering network going on between GCP and AWS, and only through that we communicate. So that's, like, the most biggest priority for us. So, yeah, that's how we handle that in today's world.
Jordan Metzner [13:42]: Okay. Cool. Yes. I understand. Like multi tenant kinda multi organization, to create kind of individual security layers. I guess maybe just last follow-up on, like, the AI stuff in regards to this, and obviously, like, through Google and Bedrock, etcetera. And obviously, like, the future providers, whatever Microsoft rolls out type in the same. But, really, I guess, like, what have these new models in their deployment inside of these platforms enabled you guys to do that wasn't able to be done, as of yesterday per se? So, whether it's, Gemini three or whether it's, Claude through Bedrock or whether it's even just like OpenAI through Bedrock or even some of the Amazon hosted models, like, are obviously really inexpensive if you run them on Titanium, if maybe they're doing things like data processing. So, yeah, it'd just be interesting to hear, as the new models have been rolled out, how have they made an impact on kind of new things you've been able to
Arun Kalaiselvan [14:33]: It's I'm I'm glad you touched on the expense of things. Our last bill was about six x for what we would expect it to be because there was a change in Anthropic and how much time it took our Lambda to actually have that connection open with Anthropic. So it was kind of insane. We're like, where did this come from? We thought it was because we launched a new customer, one of our largest ones. And then we're like, wait. No. That's not it. This was literally our AI calls to generate details, a human readable detail of a shift that is being texted to people, and this thing increased our bill six x.
Anand Chandrasekaran [15:16]: Yeah.
Arun Kalaiselvan [15:16]: We've we've been trying to fix that.
Anand Chandrasekaran [15:18]: Yeah. To touch up on that. Right? Like, of the biggest advantages that I am seeing as of today with using the latest models is we are time bound in most of the things that we do. And we have SLAs sometimes with, like, five minutes with our customers when they need a feedback. And with AI, when you do real time stuff, it takes time. Like, putting we only see a chatbot being put on, like, you're getting Claude or chatbot responding very fast. But at the same time, if you're doing, like, document processing, compliance processing, it takes a lot of time. Like, it can take multiple sometimes you have to do multi shot instead of single shot. But if you look when I run some analytics between using Claude three point like, 3.5 and what it was 1.5 to all those things, a lot of the multi shots are now single shots. That's like the biggest advantage. And also the prompt itself can be very precise. Instead of writing tons and tons of examples, you don't have to do that anymore. Whereas the structured output and all, like, it's more straightforward. That means that your token length is concise, so hallucination is less.
Anand Chandrasekaran [16:30]: And the speed at which it responds, especially Gemini three and the Gemini 2.5 flash, they are just too good. My personal thing that I'm consider currently talking with Google is the Flash-Lite. But I need that to be a little bit more capable than what it is right now. So but that's that's the kind of the fun talk that's going on right now. Right? Like that you can do a lot of things with single shot for most of the applications and that reduces the time and that has a big impact for our customers. So that's that's one of the biggest wins I'm seeing right now.
Jordan Metzner [17:04]: Yeah. That's great. Like, I'm no more, like, exclusionary kind of, like, don't do this. Don't do all those rules, like, I mean, you would just, like, run iterations and then it would do this random thing and then you'd be, like, never ever do this. And then, if it's not high enough in the ex in the exclusion rules, then, like, it'll start to do it. Right?
Arun Kalaiselvan [17:21]: That's thing, though, Jordan. We still have to do that. Because if you think about it, the restrictions that we have in place are very, very different from using a chatbot that that learns from what you're talking to. So we actually don't allow any kind of long term memory when we sent any prompts to our AI. Right? So we actually have to put in all of the exclusions, all of the context, everything in every single chat that or every single LLM query that goes out there. So, yes, a little bit less work in terms of not having to do multi shots, but at the same time, we do still have to maintain all of the context on our end. And the AI itself actually is not maintaining any kind of memory on any data that we send it.
Anand Chandrasekaran [18:07]: I think one useful tip that I would like to give to people, one of the things that I solved for our use case, it became very evident when I started experimenting with something called Pydantic AI. Pydantic is the base model in Python that's that's what was being used by Anthropic and other agents internally for their LLMs. So with Pydantic AI, when you use that, you get something called as dependencies, which is like an input type. And when you use that, almost 90 to 95% of the things that we do in our normal day to day AI agents, we don't have to give examples. You can control that using input type. And I use that quite a lot for flags, for doing things that kind of you know, it's like the best word to use is to navigate the agent's thinking in one particular way by using the control structure that it understands. After using yeah. After starting to use that, it's been extremely fast development and the hallucinations have reduced 40 to 50% for me just by changing the input types with Pydantic AI. So if you're building single agents, single short agents, I really highly recommend people using Pydantic AI for So
Jordan Metzner [19:21]: That's oh, wow. So okay. So that's like the progression from moving from like JSON type or like any type of like, formatted type prompts, I guess.
Anand Chandrasekaran [19:29]: Owned into all Pydantic classes. No more JSON types. So yeah.
Jordan Metzner [19:34]: Okay. Alright. Awesome. I'm like, I'm gonna do this next, like, after this call. After our show ends, I'm going straight to download the library and create some outputs.
Sam Nadler [19:43]: Well, I do wanna like, we're I think we're pretty deep in the weeds, and it was really insightful. I do wanna zoom out about 30,000 feet and just I know we touched upon it a little bit in your intro, but, I just wanna hear you know, I have some stats up here, and just hear about kind of the impact you're having on your customers and, the major pain points you're solving. 25% more clinical capacity. You're engaging with 60% more caregivers. Obviously, a three point five month payback is amazing. Like, what is the core solution you're providing? And it seems like it's having a huge impact on your customers so far.
Arun Kalaiselvan [20:23]: So our kind of, like, foot in the door is our scheduling ecosystem. So our scheduling agent helps manage scheduling directly for all of our customers' employees. So I'll give you an example. Let's say you're managing a bunch of nurses. Let's say you have a 100 nurses you're managing and you have schedules going on all over the place, and then this morning, one of your nurses gets sick. They say, hey. I can no longer make the rest of my shifts. One of the first things you have to do is, one, figure out which shifts that they are assigned to for that day, and then you have to find people who are able to go service or shifts that are also available and have the current credentials and are current in all of those credentials. And they like the patient likes to work with them. There's a bunch of other tribal knowledge involved as well that goes into this whole process. So suddenly, you find yourself spending two to three hours just to find the right person to fill the shift. And when you find them, you still even don't you don't even know if they're still available to do it and if they're able to do it. Right? So we take away that entire pain point for scheduling. Scheduling.
Arun Kalaiselvan [21:35]: We already know who's available, we know where they are, we know what their credential status is, we know if they're able to work with this patient or not. We literally have AI give them a phone call and say, hey, This patient's shift opened up today. They need help. We'll give you a $10.20 dollar bonus to go apply the do the shift and get it filled instantly right there. And that takes away a lot of the friction, the human friction that happens and the amount of time an admin has to sit down and take away from their regular work, mind you, while doing all of this. And that by itself has increased, like, almost 10% on average increase in revenues. That is unrealized based on existing clientele. But the even better stat is one of our customers, they sold our company to a larger entity. Their CEO now helps us expand our sales network in the in home health care space. And she was telling us about how, over a weekend, Arya filled about 12 different shifts for pediatric patients that would have otherwise gone completely unfilled if a human were managing it.
Arun Kalaiselvan [22:46]: Yeah. We are seeing impact of different kind of scales, like this is what we look for. And all of these stats you see here, they're they're pretty conservative, I would say.
Sam Nadler [22:55]: Yeah. Well, I think there's not only the time savings of the staffing agencies, also the increase in revenue, the less headaches. But, obviously, there's if there's more consistent care
Arun Kalaiselvan [23:09]: Mhmm.
Sam Nadler [23:09]: There the quality of care is just generally better. I don't even know if there's a way to I'm sure there is, but successfully measure that. But Yeah. The quality of care of the patient seems like it's gonna be better.
Arun Kalaiselvan [23:20]: Absolutely. That's that's kinda what we are hoping for as well. In fact, we had a recent shift internally on how we think about shifts. So we've stopped talking about shifts. We're talking about people now. So everything is about the patient. We don't say, hey, this shift is available. We say, this patient is available to be taken care of. Right? And that subtle change is the difference between, oh, I'm going to go service a shift versus I'm going to go take care of a patient. But also, it's a shift in thought process of how do you think about schedules, right? Instead of saying, hey, I have this schedule that's open today and I'm going to take it, we say, there's this patient who needs to be taken care of, say, every Monday of the week. And now a caregiver can go, okay, I would like to take out this patient every Monday of the week. I'm going to assign myself to, like, the rest of their care time period or whatnot. So, like, we are now building care teams rather than just filling shifts.
Arun Kalaiselvan [24:15]: And that's the big sort of the change in ideology and methodology here entirely.
Anand Chandrasekaran [24:21]: And one thing that's not spoken enough in this chapter of how we do it is we have our own, like, really trick, crude eligibility system that allows for all of this to happen. We try to get information from all of our customers, but we kind of also have our own heuristic based eligibility system that helps us get this system going. And we've been very, very successful in matching caregivers only because of that. And that is something that we built, like, one of the first things that we built when we decided that we're gonna do scheduling. And from using that, we've gotten to, like, onboard all of our clients into that because we were able to understand. I mean, Kunal, our CEO, worked with them and tried to understand what the process is and came up with a solution which we implemented. And we got it to a point where it's like you can take any customer. Most of the systems in that is going to be the same. And we kind of get the data plus add some more heuristics and that we go. We get you the results that you need.
Anand Chandrasekaran [25:20]: That's actually been one of the biggest differentiators when it comes to today, how they are seeing and how are they benefiting is is what you're seeing in this. Right? Like, the more impact is because of that set of calculations and heuristic analysis that we do and that's that's been a very, very important aspect.
Arun Kalaiselvan [25:35]: And to add a color to that, the output of that is a tiered system. Basically, we say for this patient, here is the top tier of caregivers that are eligible and these are the scores per caregiver at that pace. And then we will go from the highest scoring caregiver and, like, pull up the first tier, call them if we need to urgently fill a shift, for instance. They get the first phone calls, and then we'd go down the tiers and whatnot. So that, we're like, how do we maximize the best care possible for the patient while also maximizing returns for our customer and making sure that schedules are streamlined. Something that we're thinking about is how do you help caregivers manage their schedules in a fashion where they aren't spending a lot of time behind the wheel, for instance. So we're, like, we're looking at the traveling salesman problem where we're playing
Jordan Metzner [26:26]: Classic greedy algorithm. Right?
Arun Kalaiselvan [26:28]: Yeah. How you like how do you schedule visits for a patient where they start from home, they drive, and then the last shift is this them being pretty close.
Jordan Metzner [26:39]: You have a traveling salesman problem with time windows. I know all about this problem. Oh, man. You have no idea, guys.
Sam Nadler [26:46]: We can talk about this
Jordan Metzner [26:48]: A whole another episode.
Anand Chandrasekaran [26:50]: Keep saying that it's not a traveling salesman problem because there is a time window and known params.
Jordan Metzner [26:56]: No. I'll tell you the book is on Amazon. I can tell you the author who wrote it. And there's a bunch of traveling salesman problems you have with time windows with different types of constraints in your particular circumstance. My guess is you have like a known start time but an unknown end time of care. So you have some type of time window that you have to guarantee against. But I'd love to talk about it more offline. We could talk about traveling salesman indefinitely. Okay, let's move on to the news because this is probably maybe even getting a little too nerdy for our audience.
Sam Nadler [27:26]: Yeah. Let me kick it off. Jordan and I are no longer single. Maybe I don't know if you two are, but I thought this was an interesting article. Hinge CEO stepping down to launch Overtone, which I believe is from Match, the Match Group, to launch an AI dating app. I think it makes sense. This is eventually gonna AI is gonna, infiltrate the dating app world, but any thoughts here?
Arun Kalaiselvan [27:54]: It's funny we shouldn't we should be talking about dating because about fifteen years ago? Thirteen. Thirteen. Yeah. Thirteen years ago, Anand and I, we lived in New York. Went to NYU together. We worked basically at the same companies for a long time. We go way back. And one of our hackathon projects was a we call it e flirt engines. It was a matching engine for you would come in and answer a bunch of questions. You need to be on these dating services. So we actually know all of these folks pretty well. I think this is an interesting move, especially with the timing. On one hand, you have Mamdani who's all over the news as having met his wife on Hinge. Right? And, like, so Hinge is definitely getting a little push on the PR end from that side. But then this whole situation where oh, man. They I it was inevitable that this happens. I am surprised that it was match.com that started to overtone and not someone else. I would have assumed the folks at Bumble would be the ones doing it, but sounds like they are doing they have their own version of this coming as well.
Arun Kalaiselvan [29:05]: I think Justin is probably one of the best people to push this forward too, so this is a fantastic move on Match's part. What I don't know yet is how this is going to devolve into just matchmaking versus AI dating. Like, the are we having our AI personas date each other? And, like, how does that track to real world entities? And then how do you go from setting things up with AI to the real world? I think that's going to be the Bayesian. Maybe we don't. I don't know.
Anand Chandrasekaran [29:38]: Yeah. I think it's for me, right, like how the when AI was released by Google for photos, the question was, what does it mean to even take a photo? Right? That question is we if you go into Reddit, we still talk about it a lot. Same way, like, if people when they write their profile in a dating app, my single most important question is how true is that? And so is it how are they going to solve that problem? I mean, already, a lot of it is copy paste. But with AI, it is just the best tools to do that to make it sound nice. It might help a lot of people, but I'm I'm curious to see the interaction. I'm I was always sure that this was going to happen, so I'm not surprised. I'm I'm curious to see how this will work.
Arun Kalaiselvan [30:28]: I think this will be fantastic for the first meet, but AI only knows about your highlights. Right? Let's let's say if it says, oh, I'm gonna look at your Facebook and see everything that you've done and then you'll make your profile using that. That's all the highlights. But then, eventually, it's it's the midlights that matter for truly good relationships. And I think the people who will win are the ones who are able to somehow focus on that rather than just all of the highlights of a person.
Sam Nadler [30:58]: So it's not gonna capture that you snore and that you Right? Yeah.
Arun Kalaiselvan [31:03]: Like, you have a CPAP machine. Yeah. Exactly.
Jordan Metzner [31:06]: Alright. Let's move on.
Sam Nadler [31:08]: Alright. Alright. Now, I'll kick you off with you. Cursor CEO believes OpenAI, Anthropic competition won't crush his startup. I know you're a heavy Cursor user, and but you guys also have some thoughts on, how it's it's position as the application layer versus the model layer.
Jordan Metzner [31:30]: I mean, have every IDE installed. Like, I have Visual Studio. I have Antigravity. I have Windsurf. I have Cursor. My go to is cursor. It's like I think it's definitely the most developed. And when I can use, like, multiple agents simultaneously, it's definitely the best, like, product user customer experience. I've been using Antigravity a little bit, but, it's free, so I use try to use up my tokens until they kick me off. But it's not as polished as a product. It does have some cool, like, one off features that are pretty cool. So I thought, like, obviously, there's just so much copying going on or, like, I don't know, like, feature, like, feature parity might be, like, the proper word. So, like, if anybody launches anything, everyone else will just launch it. If it's cool enough and, like, the time to market is so fast. Like, and my Cursor, I can't even use it for one day without requiring, like, a feature up you know, a product update on it.
Jordan Metzner [32:17]: So I think at the end of the day and honestly, it is surprising to me how good Claude Code is and, like, the path that Cursor has gone, the path that Anthropic has gone because like I think that Anthropic could like, choose a selfish model where they would literally sell their, code editor only in their tools and nowhere else. And I don't know. To me, it would seem like an even bigger opportunity. Maybe I don't like, I'm sure they've done that evaluation. Dario's a pretty smart guy, but, sell to everyone has been their model. But, at the end of the day, it's whatever the best coding model is. In my personal opinion, I try to think of costs not being a factor because my coding time is so expensive to me. Right? We have you know, these Anthropic models are so expensive. We have seen that they cost more money. Like, I see my cursor credit card getting hit all the time. But, realistically, like, I don't know. Does this, like, essentially see the inflationary values here, where this just becomes cheaper. I think so because already I'm running four agents simultaneously when, like, before I wasn't. So, I think we'll see that.
Jordan Metzner [33:23]: But yeah. I don't know. What do you guys think? How are you guys coding? And, like, how do think?
Arun Kalaiselvan [33:26]: We are controversial on Cursor.
Anand Chandrasekaran [33:28]: I mean Okay. Cool. We like Cursor. It's it's a good product. But this when I read this article, I kind of understand where she's coming from. Right? Like, so think about it like this. Two years ago, Claude was the only coding model. Today, we can actually do some work with Gemini three. Still not yet there. But in two years from now, I'm go there is going to be a saturated simple state for coding from most of the models. So at that point, you are no longer going to be dependent on Claude. And at that point, it was going to become tooling is going to be the most important aspect. And then it will be like, is going to be giving you all those things at the less the least cost. And that time, Cursor will be probably the top in the market line because they've already been trying to do the best tooling and the way the interactions and all for a long time, and they would be in a very matured state, and they will all be able to offer things for its much cheaper price.
Anand Chandrasekaran [34:28]: I don't know where Claude Code will be at that point, but Claude Code is already very expensive when compared to everything else. And then I'm I'm assuming there's going to be some enterprise system play going on by Cursor where it's it's like it's going to be adapted by many people. So I do think that it's going to what he's saying is right and if that's the way it would go because there is a huge competition to build the best coding model still exist. And once it's solved, like, we're still fighting with Claude and everybody's compete. I don't know long
Arun Kalaiselvan [35:01]: You can solve that, though. There is no clear definition of best coding model. Like, there is no clear definition of best engineer. Right? Like, it's a if you can get there to the ninety fourth percentile correct, then that's all that really matters.
Anand Chandrasekaran [35:14]: That's what I'm saying. Right? That's that's Okay. But
Jordan Metzner [35:16]: But let me push back on you guys for a sec. Yeah. If you say that, like, talking, as we spoke earlier, like, using this Pydantic model, like, talking to the models, if you know how the model is built, the more you the closer you talk to how the model is built, the better are you at like getting the model's response rate. Then you would have to argue that those who build the model would be best at creating the tooling around said model. So while you well, let me just like finish this thought. Right? While that might be so might be true, so if inherently, like if all these models to your point kind of reaches logarithmic curve, then he who has one of these top models and is able to best tooling around their set model and cost might be the number one definitive thing in the in that sense. Right? Because like, hey, do you want an AI code developer that's doing it at, a dollar an hour or at $200 an hour or at 50¢ an hour or at 5¢ an hour and like, does, like, going all the way back to, like, the chips, like, is Google Silicon able to, like, then allow you to use Gemini at this logarithmic curve and they built the best tooling around it and then their IDE because, like, you have to understand there the Google bought, this windsurf. Right?
Jordan Metzner [36:23]: Got You've Microsoft that has VS Code that's already open source. Right? We haven't seen anything from Amazon but, like, maybe they would buy it. Maybe they would buy cursor. Maybe it looks like cursor is almost too big and it's gonna run alone. So, is the cross stack tooling across all these things important? If we reach this log driven curve and it doesn't matter and all the models are the same, then you just want the cheapest best one. And so then he who builds the model may be the owner of that and cross tooling might not be important at all. I don't know. I'm just trying to flip the other side of it.
Arun Kalaiselvan [36:49]: But here's the deal. At that point, is tooling even important? If you have a AI coding system that's doing it, why do you even use need to use the tools? So for us with Cursor, we did a whole six month pilot with Cursor. We were on their enterprise ecosystem. We actually found that they introduced more issues into our code base than they solved for. So we ended up completely shutting our cluster from our eco Cursor from our ecosystem, and we switched to a Copilot.
Anand Chandrasekaran [37:15]: Yeah. I think that was back in the day when they were trying to figure things out.
Arun Kalaiselvan [37:19]: This is about, what, two years ago at this point?
Anand Chandrasekaran [37:21]: One and a half years ago. Yeah. But Cursor is now very mature. I mean, yeah, I see. Like I said, I've actually had a chance through a friend to directly talk with one of their co founders or CEOs. I'm not sure. But, like, we had discussions and I gave them reports and stuff. But my point is to talk about what you are saying. Yeah. But you're right. With if Claude becomes the winner of that, then yeah. But more often than not, like in the technological world, there is always an alternate in anything that you do. And I don't believe that there is going to be one single there is always going to be two or three. Yeah. And there's always going to be this open source thing. And like, if we but look at Andrej Karpathy doing things. Right? He's like built he built his own new LLM model that you can just run, and then now he has built this judge concept that that that you can do use for anything. Right? So there are gonna be people who are gonna be pushing open source stuff all the time.
Anand Chandrasekaran [38:14]: In the last ten years, the most successful code editor has been VS Code from Microsoft, which was which has been enterprise for the longest time, embraced open source, and they became that. So that's that's how I'm I'm I'm seeing it.
Arun Kalaiselvan [38:26]: So as far as evolutions go, yes. What I don't know is what the next evolution looks like. And we went from I was a religious VIM user. I went to Sublime and then from Sublime, we went to Visual Studio. Right? So it's a and of course, cursor is is basically VS Code. Yeah. It's just a
Anand Chandrasekaran [38:47]: Yeah. Is VS Code. Yeah. Antigravity is VS Code.
Jordan Metzner [38:50]: Yeah. Of course. That's why I'm so long Google because I think, they can have a top model provider for cheap and, they got the ID
Arun Kalaiselvan [38:56]: I do think Google is going to be the silent killer in this current ecosystem. The way we've seen the Vertex AI team operate in the last three years has been fantastic. I think Amazon has a lot of catching up to do if they want to be competitive. They I've I've seen this in my blockchain days as well. Right? Like, all the clouds got it wrong for the blockchain world. And they've at least done a better job on the AI side of adoption, but Google is, like, leaps and bounds beyond where Amazon is. I can't really comment about where Microsoft is right now, but we are definitely all in on the Gemini team.
Jordan Metzner [39:37]: Yeah. That's awesome. Alright. Great episode, guys.
Sam Nadler [39:40]: Yeah. No. I just wanna say thank you so much. Thanks for joining. And, before we cut, just like and subscribe. We're here every Friday. Thank you, Anand. Thank you, Arun. And please check out Arya Health. Thank you so much for having us.
[39:57] Theme song: Build this week, breaking it down
